-
Notifications
You must be signed in to change notification settings - Fork 6
/
train_ae.py
80 lines (58 loc) · 2.58 KB
/
train_ae.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
from __future__ import print_function
import torch
import torch.utils.data
from torch import nn, optim
from torch.nn import functional as F
import numpy as np
from .dataloader_ae import dataloader
from models.ae import motion_ae
torch.backends.cudnn.benchmark = True
def train(model):
optimizer = optim.Adam(model.parameters(), lr=1e-3)
tf = 1
gradient_clip = 0.1
max_iter = 80001
model.train()
for batch_idx in range(max_iter):
seq_len = np.random.randint(3, 100)
data, wh = next(iter(train_loader.generate(seq_len)))
data = data.cuda()
data_vel = data[:, 1:, :] - data[:, :-1, :]
optimizer.zero_grad()
reconstructed = model(data_vel, tf)
loss = F.mse_loss(reconstructed, data_vel, reduction="sum")
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), gradient_clip)
optimizer.step()
if batch_idx % 50 == 0:
tf *= 0.99
model.eval()
with torch.no_grad():
test_sample, wh = next(iter(val_loader.generate(100)))
test_sample = test_sample.cuda()
test_sample_vel = test_sample[:, 1:, :] - test_sample[:, :-1, :]
reconstructed = model(test_sample_vel, 0)
np.save("temp/ae/train/reconstructed.npy", reconstructed.cpu().detach().numpy()[0])
np.save("temp/ae/train/original.npy", test_sample.cpu().detach().numpy()[0])
np.save("temp/ae/train/wh.npy", wh.numpy()[0])
print("[", batch_idx, "]\tLoss: ", round(loss.item() / seq_len, 4), "\ttf: ", round(tf, 4))
model.train()
if batch_idx % 10000 == 0:
torch.save(model.state_dict(), "checkpoint/ae/ae_" + str(batch_idx//10000) + ".pth")
def evaluate(model):
with torch.no_grad():
test_sample, wh = next(iter(val_loader.generate(100)))
test_sample = test_sample.cuda()
test_sample_vel = test_sample[:, 1:, :] - test_sample[:, :-1, :]
reconstructed = model(test_sample_vel, 0)
np.save("temp/ae/eval_reconstructed.npy", reconstructed.cpu().detach().numpy()[0])
np.save("temp/ae/eval_original.npy", test_sample.cpu().detach().numpy()[0])
np.save("temp/ae/eval_wh.npy", wh.numpy()[0])
if __name__ == "__main__":
train_loader = dataloader(split="train", batch_size=512)
val_loader = dataloader(split="val", batch_size=1)
model = motion_ae(256).cuda()
train(model)
#model.load_state_dict(torch.load("checkpoints/ae_weights/ae_combined_path_mot_8.pth"))
#model.eval()
#evaluate(model)